Therefore, with or without the histone modification information, the predictive model established on the CD4 lymphocyte data can well generalize to the other tissue or cell type data.
For generalizability testing on M1, we apply the histone modification information of the CD4 lymphocyte to the other 11 tissues and cell types because correlation analysis by ourselves and others has indicated that histone modifications exhibit modest to strong correlations for different cell lines [ 39, 13].
However, if compared to the validation results where the histone modification information is not used (see Table 3), the performance on the testing data is not unexpected.
The two predictive models, one with the histone modification information (M1) and the other without (M2), that are both built on the human CD4 lymphocyte data are then tested on the data of the other 11 tissue and cell types for their generalizability.
When the histone modification information is not used, the performance of the predictive model on the data of the other tissues and cell types deteriorate substantially, especially in terms of the sensitivity.
The sensitivity, specificity, accuracy and correlation coefficient of M1 and M2 during these testing experiments are summarized in Tables 6 and 7. When the histone modification information is incorporated, the classifier model built on the CD4 lymphocyte data can be applied to most of the other tissues and cell types (except for sperm) with little or no performance deterioration.
Note that both models have incorporated the histone modification information.
Mikkelsen et al. [ 28] made the mouse histone modification information available and the human histone modification information was originally from Pan et al. [ 37].
Through these experiments, we assess the individual and combinational influence of the newly added features and the impact of histone modification information.
The modification detector assigns modification information (i.e. negation and speculation) to each event.
